Churn Prediction Without a Data Team: A 4-Step Weekly Model

When a Head of CS at a 90-person SaaS told me she needed "an ML model to predict churn," I asked what she'd do with the score on Monday morning. She paused. That pause is why most SMB churn-prediction projects die — they're built before anyone defined the Monday action.

Why do most SMB churn models never ship?

Because they get scoped as data-science projects instead of CS operations. The team chases a logistic regression on 18 months of cohort data, hires a contractor, and 14 weeks later still doesn't have a Monday workflow. Meanwhile two accounts churned quietly.

Definition: Churn signal — a measurable change in account behavior that statistically correlates with cancellation within 30-90 days.

The honest version: a Head of CS who knows the book of business can mentally rank the at-risk accounts in 20 minutes. The job of the model isn't to replace that judgment — it's to make sure no account gets missed when the CS team grows past what one person can hold in their head.

What does a 4-step model actually look like?

Four steps, weekly, no ML infrastructure. Each step is a specific deliverable a CS lead can run on a spreadsheet plus an LLM.

Step 1 — List the 9 behavioral signals

Not 50. Nine. The signals split into three groups: usage drift, relationship drift, support drift.

Usage drift (3 signals):

• Logins per active seat down 30%+ vs 4-week average
• Core-feature usage down 40%+ for the workflow you sold them
• New-user provisioning stopped (no seat additions in 60 days when prior pattern was monthly)

Relationship drift (3 signals):

• Champion left the company (role change, LinkedIn signal, or out-of-office that becomes permanent)
• Exec sponsor stopped attending QBRs or last 2 meetings rescheduled
• Renewal owner on client side changed and didn't reach out

Support drift (3 signals):

• Ticket volume up 50%+ and sentiment turned negative
• Repeat tickets on the same root cause (3+ in 30 days)
• Last NPS or open-text feedback contained a comparison to a competitor

These 9 cover ~80% of churn cases I see in B2B SMB. You don't need more until you've worked these for 6 months.

Step 2 — Score each account weekly

A flat-weighted score, one column per signal, 1 if present, 0 if not. Sum out of 9. Tier:

• 0-2: green (normal monitoring)
• 3-5: yellow (proactive outreach this week)
• 6-9: red (executive escalation, save play)

Definition: Save play — a defined sequence of actions a CS team runs against red-tier accounts: discovery call, exec sponsor activation, success plan rewrite, and a documented decision point at day 30.

Resist the urge to weight signals differently before you have data. Equal weighting beats fake precision for the first 3-6 months. After 25-50 churn events, look at which signals fired before each loss and re-weight.

Step 3 — Use AI to write the account-level summary

This is the step that actually changes the workload. For each yellow/red account, an LLM reads:

• Last 30 days of usage telemetry (a CSV export)
• Last 90 days of support tickets (subject + first message)
• Last 4 weeks of meeting notes (if the CS team takes them)
• NPS open-text comments from the last quarter

And produces a 6-line narrative: what's drifting, who's involved, what the likely root cause is, and the one question to ask in the next outreach. That narrative is what the CSM walks into the save call with.

Step 4 — The Monday 30-minute review

The Head of CS plus 1-2 senior CSMs spend 30 minutes Monday morning on red and high-yellow accounts. Each account gets a named owner, a save-play assignment, and a check-in date. The list goes back into the spreadsheet with last week's actions marked done/not-done.

That's the loop. Sunday night the AI generates the narratives, Monday morning humans decide. Build does not exceed 4 hours; weekly cost is 90-120 minutes of CS time plus whatever the LLM calls cost (typically under €50/month for an SMB book).

Copy/paste weekly tracking template

One row per account in the yellow or red tier. New tab per week.

Week of: [DATE]
Account: [NAME]
ARR: [N]
Renewal date: [DATE]

Signals fired (1/0):
- Logins down 30%+: [ ]
- Core-feature use down 40%+: [ ]
- Seat additions stopped: [ ]
- Champion gone: [ ]
- Exec sponsor disengaged: [ ]
- Renewal owner changed: [ ]
- Ticket volume +50% with bad sentiment: [ ]
- Repeat tickets (3+ in 30d): [ ]
- Competitor mention in feedback: [ ]
Score: [N]/9   Tier: [G/Y/R]

AI summary (6 lines):
- Drift: [TEXT]
- People: [TEXT]
- Likely cause: [TEXT]
- Open question: [TEXT]
- Prior save plays: [TEXT]
- Risk to ARR if lost: [N]

Owner: [NAME]
Save play: [TEMPLATE_KEY]
Next check-in: [DATE]
Action this week: [TEXT]

The template is deliberately boring. Boring survives.

Tool tip (AIAdvisoryBoard.me): The pattern under this is Plan → Fact → Gap applied to the customer base: the renewal plan you sold (Plan), the actual behavior the account is showing (Fact), and the early-warning gap between them (Gap). Most SMB CS teams have Plan and Fact in separate tools and never compute the Gap until renewal week — which is too late. A daily-management OS surfaces the Plan → Fact → Gap automatically across accounts so the Monday review starts with a pre-ranked list, not a manual scan. See how it works at https://aiadvisoryboard.me/?lang=en.

Manager scan (2-minute digest example)

• Plan: every account in our book has a defined renewal path (date, ARR, success criteria) — Fact: 14 of 47 accounts have no documented success criteria — Gap: those 14 are flying blind regardless of score
• Plan: at-risk accounts get a save play started within 5 business days of turning red — Fact: 3 of last quarter's reds waited 11+ days — Gap: Monday review is happening but follow-through isn't
• Plan: champion-loss signal triggers a within-week outreach — Fact: 2 champions left in March and got no outreach for 6 weeks — Gap: LinkedIn change-of-job alerts aren't wired in
• Plan: every QBR rescheduled twice triggers exec sponsor activation — Fact: pattern exists but isn't acted on — Gap: nobody owns the "twice-rescheduled" trigger
• Plan: AI narrative generated before Monday review — Fact: running on time 11 of 13 weeks — Gap: minor, monitor
• Plan: signal weighting reviewed after each churn — Fact: 4 churns this quarter, no post-mortem in 3 of them — Gap: the loop isn't closing

Micro-case (what changes after 7-14 days)

A 140-person B2B SaaS deployed this exact 4-step model after losing two six-figure accounts in one quarter — both of which "had no warning signs" according to their CSMs. Week 1: spreadsheet built, signals defined, baseline pulled. Week 2: AI narratives running, first Monday review held — and immediately flagged 4 accounts the CS team had not been actively worrying about, including one where the champion had left 3 weeks earlier and nobody had noticed. By week 6: red-tier accounts down from 9 to 4, two saves closed (one with an expansion attached after the discovery call surfaced a different pain point), and the Head of CS got back roughly half a day per week she used to spend scanning the book manually. Cost: about €40/month in LLM calls. No data team.

Note on this case: This example is illustrative — based on typical patterns we observe with companies of 30-500 employees, not a single named client. Specific numbers are rounded approximations of common ranges, not guarantees.

Tool tip (AIAdvisoryBoard.me): The Monday review only works if the signals are computed on time — and most SMB CS teams discover that their telemetry, CRM, and support data live in different systems with no nightly join. The Plan → Fact → Gap layer in our daily-management OS handles the join for you: nightly pull from the source systems, signal scoring against the Plan you defined, ready by 7am Monday with the at-risk list ranked. The 7-day diagnostic shows whether your current data shape can support the 9-signal model or needs one cleanup pass first. Start at https://aiadvisoryboard.me/?lang=en.

FAQ

What if we don't have usage telemetry at all? Then start with relationship + support drift only (6 of the 9 signals). That still catches the majority of B2B SMB churn — relationship signals are stronger than people expect. Add usage telemetry on the next product release; don't wait for it to start the model.

Won't equal-weighted scoring miss high-impact signals? For the first quarter, no — equal weighting beats overfit weighting on small data. Once you have 25+ churn events, look at which signals fired in the 60 days before each loss and re-weight. Before that, weighting is decoration.

Should the AI also draft the outreach email? Sometimes useful, but the higher-leverage use is the account narrative — the email is the easy part once you know the story. Get the narrative right first; the email is one prompt away.

How is this different from a customer health score? A health score is a single number; this model is a structured workflow. Health scores often get computed and ignored. The 4-step loop forces a Monday decision — which is the only thing that prevents silent churn.

What about expansion signals — do they belong in the same model? No, keep them separate. Mixing risk and expansion signals creates a score that means nothing in either direction. Run a parallel expansion model on the same cadence; the 30-minute Monday review covers both.

Conclusion

A 4-step churn model with 9 signals and an AI narrative beats a 14-week ML build for any SMB that does not already have a CS team larger than ten people. The point isn't sophistication; it's that no account gets missed, every red gets a named owner, and the loop closes weekly.

Pick the 9 signals. Build the spreadsheet this week. Run your first Monday review next Monday.

If you want a system that surfaces the Plan → Fact → Gap automatically — every day, across the customer base — see how the 7-day diagnostic works at https://aiadvisoryboard.me/?lang=en.